Metric Learning with Feature Embedding for Segmentation Quality Evaluation

Huixiang Chen, Bo Peng, Zaid Al-Huda, Xiaofan Li

Abstract: Segmentation quality evaluation is an essential step to quantity the performance of segmentation algorithms. It can be used as a feedback for correcting segmentation errors or selecting appropriate algorithm parameters. We propose a novel evaluation framework where a convolutional neural network is designed for distinguishing the segmentation quality instead of using ground truth images. Our work has three primary contributions: First, we evaluate the quality of object segmentation by learning region features. A novel feature embedding model is proposed to integrate meta evaluation principles in a metric learning process. Second, it exempts the requirement of ground truths in the test stage, where object features of trained classes are used for the discrepancy calculation. Third, a large-scale object segmentation evaluation dataset is constructed, which contains various segmentation qualities under different assumptions. The experimental results on PASCAL VOC2012 dataset demonstrate that our method improves the evaluation accuracy and outperforms the popular supervised evaluation measures.